Comparative Study of Combination of Preprocessing, N-Gram Feature Extraction, Feature Selection, and Classification Method in Indonesian Sentiment Analysis with Imbalanced Data

Abstract

Social media makes a shift in lifestyles of people. People tend to use microblogging such as Twitter to criticize the controversial issues. The most controversial Indonesian economics policy in recent year is tax amnesty. Predicting positive and negative sentiments on tax amnesty policy could be developed by supervised machine learning. The performance of classification results can be improved by using the right combination of preprocessing technique, feature extraction, and feature selection. We aim to compare the performance and find the best combination of preprocessing technique, N-Gram feature extraction, feature selection, and classification method by conducting experiments. Data collection was developed by crawling using Twitter API. Imbalanced data is one of challenge in machine learning which can produce unsatisfactory classifiers and normalize the Indonesian slang can also be more challenging. This research uses an imbalanced dataset to know the performance of the combination algorithms handling the imbalanced data which measured by nested cross-validation. The experimental results show that the best combination of algorithms in this research performs well in handling imbalanced data and the performance of models can be improved and really depend on the combination of preprocessing, N-Gram feature extraction, feature selection, and classification method.